计算机科学
机器学习
人工智能
药物发现
有向无环图
深度学习
极表面积
生物利用度
生化工程
虚拟筛选
广告
图形
先验概率
铅(地质)
强化学习
作者
Davide Boldini,Jesko Kaiser,Heide Marika Duevel,Christin Rakers
标识
DOI:10.26434/chemrxiv.15004114/v1
摘要
PROTACs (Proteolysis Targeting Chimeras) are bifunctional degraders that harness the ubiquitin–proteasome system to eliminate disease-relevant proteins. Their large molecular size and inherent beyond-Rule-of-Five properties pose significant challenges to oral bioavailability, and reliable predictive models for intestinal absorption remain elusive. Here, we present a deep learning framework for PROTAC oral bioavailability prediction based on foundation model finetuning of a graph neural network pretrained on diverse ADME endpoints. A key innovation is physics-based multitask finetuning, in which expert-knowledge descriptors including effective polar surface area (ePSA), chromatographic LogD, and quantum mechanically derived radius of gyration and solvent-accessible surface area (SASA) serve as auxiliary tasks, embedding structural priors into the model without compromising virtual screening speed. Benchmarked against score-based approaches (ETR, BBI, AB-MPS) and classical machine learning baselines across cross-validation, temporal, and project-based splits on a dataset of over 200 in-house mouse pharmacokinetic studies, the finetuned foundation model consistently outperformed all alternatives in interpolation and temporal extrapolation settings. Prospective application in active drug discovery projects demonstrated its utility for prioritizing compounds balancing cellular potency with favorable predicted absorption.
科研通智能强力驱动
Strongly Powered by AbleSci AI